An Analysis of the Heterogeneity and IP Packet
Reordering over Multiple Wireless Networks
Dominik Kaspar∗ , Kristian Evensen∗ , Audun F. Hansen∗ , Paal Engelstad†∗ , Pål Halvorsen∗‡ , Carsten Griwodz∗‡
∗ Simula
Research Laboratory, Norway † Telenor R&I, Norway ‡ University of Oslo, Norway
email:{kaspar, kristrev, audunh, paalee, paalh, griff}@simula.no
Abstract—With the increasing deployment of wireless technologies, such as WLAN, HSDPA, and WiMAX, it is often the case
that simultaneous coverage of several access networks is available
to a single user device. In addition, devices are also often equipped
with multiple network interfaces. Thus, if we can exploit all
available network interfaces at the same time, we can obtain
advantages like the aggregation of bandwidth and increased fault
tolerance. However, the heterogeneity and dynamics of the links
also introduce challenges. Due to different link delays, sending
packets of the same flow over multiple heterogeneous paths causes
the reordering of packets.
In this paper, we quantify the impact of network heterogeneity
and the use of multiple links on IP packet reordering. We
show with practical measurements, according to commonly used
metrics, that packet reordering over multiple links exceeds the
reordering caused by common connections in high-speed, widearea networks. We also demonstrate that heterogeneity and
reordering exceed the assumptions presented in related work.
By using sufficiently large buffers, packet reordering can be
avoided. However, for devices with high resource constraints,
the workload of using large buffers is expensive. Sender-side
solutions of dividing and scheduling a packet sequence over
multiple links can reduce the buffer requirements at the receiver.
Initial experiments with a static scheduler, that has knowledge of
average link delay and throughput estimates, show that packet
reordering can be reduced by only 38 % due to the dynamic
heterogeneity of the two links.
Due to the currently existing challenges, end users do not
yet have easy access to exploiting multiple networks at once.
However, with modifications of an operating system’s routing
tables, it is possible to achieve a multilink configuration, which
is able to achieve aggregated throughput, provided that an
application is used that opens many concurrent connections.
A BitTorrent client is a perfect example of such an application, because the underlying protocol logically splits a
large file into segments, which are downloaded independently
and concurrently from various locations. A new connection
is opened to each provider of a data segment, and using a
multilink setup, this can be exploited without any changes
to the protocol. After defining an equal-cost multipath route
(ECMP) [2] for each available interface, each new connection
will automatically be opened over one of the interfaces. For a
large number of connections, traffic will be equally balanced
according to the bandwidths of the links. Figure 1 illustrates
the benefit of a multilink setup using an HSDPA and a WLAN
link aggregated through the use of ECMP.
I. I NTRODUCTION
A growing infrastructure of heterogeneous wireless technologies, such as WLAN, HSDPA, and WiMAX, often places
a single user device into the coverage areas of multiple access networks simultaneously. Currently, from an application’s
point of view, Internet connectivity is usually provided using
a single link at each point in time. It is, however, increasingly
common that devices such as laptops, phones, and PDAs are
equipped with multiple wireless interfaces. Motivated by these
trends, wireless providers are looking for solutions that can
fully utilize multiple technologies concurrently when present.
Exploiting multiple links simultaneously has several benefits: increased throughput by bandwidth aggregation, added
fault tolerance by sending redundant data over independent
links, and increased mobility and connectivity by overlapping
several coverage areas [1]. However, the deployment of a
readily usable multilink solution has so far been hindered
by substantial network heterogeneity. High variances in the
characteristics of wireless links cause packets to be severely
reordered, which greatly impairs the performance of transport
layer congestion control mechanisms.
Fig. 1. Throughput aggregation with BitTorrent over simultaneous HSDPA
and WLAN links (results were obtained by downloading a total of 75 copies
of a 700 MB large file from a popular source).
These results show the potential of utilizing multiple links
simultaneously. However, building this functionality per application is not a viable approach. From an Internet service
provider’s (ISP’s) point of view, we envision the users to
access any type of services on the Internet using multiple
wireless links. For such an approach to be deployed, we
think that a solution must be able to split a single transport
connection over the alternative links. This can be solved by a
proxy or server implementation. The latter requires that, for
instance, the ISP controls the server.
II. R ELATED W ORK
For several years, contributions to multilink transfer have
been proposed on many layers of the protocol stack, ranging
from network level multihoming [3], over multipath transport
for homogeneous and independent wired paths [4], to methods
for application layer striping [5]. A good overview of recent
solutions to transport layer striping is given in [6].
Most of the related literature focuses on multipath routing
in wired, high-speed networks. In addition, the majority of
existing solutions to network striping assume substantial protocol and end-host modifications, which may hinder general
deployment. The fact that most of them have only been tested
in simulations, often based on very simple assumptions about
heterogeneity, enforces our skepticism on deployability [7],
[8]. Our field measurement results (see Section III) contradict
frequently made assumptions about the characteristics of heterogeneous links, which are often modeled too evenly and not
very realistically (e.g., 30 milliseconds delay for both WLAN
and UMTS [7]).
The effects of multipath routing on packet reordering
have been addressed in various publications ranging from
simulation-based analyses [9], over real-time measurements
in wide-area networks [10], [11], [12], [13], to IETF standardization efforts that define metrics for measuring packet
reordering [14], [15]. However, no publications have come to
our attention that address how severely packets get disordered
when using multiple interfaces.
III. M EASUREMENTS ON L INK H ETEROGENEITY
Having access to the Internet over multiple links at the same
time has many potential benefits, but it also introduces new
difficulties. The main challenge addressed in this work is that
splitting a packet sequence over heterogeneous IP paths will
most likely cause the sequence to be out of order at the destination. Especially in wireless networks, the amount of reordering
is heavily dependent on a large number of parameters that are
beyond the user’s control, such as unpredictable interference
and varying signal quality, other users contending on the same
channel, overloaded routers, etc.
Link heterogeneity has several aspects, of which the transmission delay is the most crucial factor for causing packets
split over different paths to arrive out of sequence. Table I
shows the round-trip delay of ping packets measured through
an entire day over different wireless technologies. From these
numbers, we expect that packet reordering over multiple links
will be far beyond the 0.01 % to 1.65 % level [11] stated in
the literature.
TABLE I
ROUND - TRIP TIME MEASUREMENTS OVER HSDPA AND WLAN
min.
max.
mean
stdev.
loss
WLAN1 2.2 ms 5010 ms 10.1 ms 61.2 ms 2.7 %
WLAN2 3.2 ms 1807 ms 11.2 ms 59.0 ms 1.5 %
HSDPA 67.5 ms 1350 ms 212 ms 106 ms 0.8 %
A solution to multilink transfer must also consider that link
characteristics may vary significantly over time. An example
of variability over time is illustrated in Fig. 2, which shows
the throughput dynamics of our HSDPA link.
2.8
HSDPA
Midnights
TCP Throughput (Mbit/s)
Real-time and streaming applications (e.g., video conferencing) serve as other examples of why simply dividing the
traffic over multiple links, using a vast number of connections,
is not a viable approach. The packets in such streams should be
delivered in the order of application data playout, not risking
a long pause due to a delayed bulk of packets on a slow link.
The problem of packet reordering may be mitigated by using
buffers that hold at least as much data as is “in flight” at all
times over all interfaces. However, for low-performance and
battery-operated devices, such as cell phones and PDAs, large
buffers are very resource-intensive. Keeping in mind the goal
of reducing the workload of such devices, this paper studies the
fundamental characteristics of packet reordering when using
multiple links. We also evaluate the potential gain of a static,
sender-side packet scheduler that tries to minimize reordering
at the client by sending packets ahead of time over links with
high delay.
This paper’s main contribution is to shed some light into the
effects of splitting a packet sequence over multiple wireless
interfaces and how the dynamic heterogeneity of these links
cause reordering at the destination. In Section II, currently
proposed approaches to data transfer over multiple links are
presented. Based on measurements of the heterogeneity of
wireless links, Section III then indicates the dynamics of
wireless links and points out the challenges with link aggregation. In Section IV, metrics for the quantification of packet
reordering are introduced, before Section V experimentally
examines the effects of using multiple links on IP packet
lateness. Methods of reducing packet reordering are discussed
in Section VI, and the conclusions of this work are finally
drawn in Section VII.
2.4
2.0
1.6
1.2
0.8
1
2
3
4
5
6
7
8
Time Elapsed (days)
9
10
11
12
13
14
Fig. 2. HSDPA throughput varying over a period of 14 days. The results
were obtained by repeatedly downloading a 5 MB large file over HTTP.
At present, no complete and readily deployable multilink
solution exists due to implementation challenges and substantial network heterogeneity. It is a great challenge to deal with
transport layer packet reordering and to optimally schedule
packets (or sessions) to available links that vary significantly
in their characteristics.
IV. PACKET R EORDERING
The Internet Protocol has no mechanisms to ensure that
packets arrive at a destination in the same sequence that
they were sent out at the source. There are various reasons
for the occurrence of IP packet reordering; traditional causes
include link layer retransmissions and priority scheduling in
routers. When dividing IP traffic over heterogeneous paths,
however, much higher reordering can be expected. This section
describes our testbed for conducting field measurements for
packet reordering and presents the metrics used in the evaluations.
A. Testbed Setup
For measuring the effects of packet reordering, a dedicated
server with a single 100 Mbit/s Ethernet network interface
was used. As illustrated in Figure 3, a client laptop that is
connected to both a WLAN and an HSDPA access network
receives a packet sequence from the server. In the initial
connection request, the client informs the server about its
available interface addresses and about all other parameters
used for the intended test run (i.e., bitrate, duration, packet
size, type of scheduling, etc.). When the server has received
all necessary information, it sends UDP packets at the specified
constant bitrate to the client addresses. A scheduler at the
server decides for each packet to which client address it will
be sent. If not otherwise specified, round-robin scheduling is
used. A static scheduler based on predetermined average link
delays will be discussed in Section VI-B.
the reordering. Most of the current research discussing packet
reordering is focused on single-path flows through wired, highspeed networks [12]. The common portion of packets affected
by reordering is in the order of 0.01 % to 1.65 % [11], or 1 %
to 1.5 % [10].
However, MT P RRS does not contain any information about
the severity of disorder in a sequence. In an extreme case of
receiving a packet sequence, such as S = {2, 1, 4, 3, 6, 5, ...},
where pairs of sequence numbers are “flipped”, MT P RRS
would be equal to 0.5, indicating that 50 % of all packets
arrived late. In our testbed, where two links induce different
delays, MT P RRS is therefore naturally expected to approach
0.5. Thus, for the case of splitting a stream over heterogeneous
links, a more sophisticated metric that is able to convey the
total magnitude of disorder in the received sequence would be
valuable.
One such metric, called Reorder Density, is introduced in
[13] and RFC 5236 [15]. To compute the Reorder Density
MRD , the sequence number of each packet is compared to
the position in the sequence received at the destination. For
each packet, this allows the calculation of a displacement
within the sequence, negative values indicating “early” packets
and positive displacements indicating “late” packets. MRD is
calculated by counting the number of packets with a specific
displacement. Figure 4 shows a measured example of MRD ,
obtained by sending packets over NW LAN1 and NHSDP A at
2.0 Mbit/s. The reason why the depicted MRD histogram is
not symmetric is that 2.1 % packets were lost over the WLAN
link, while the HSDPA link suffered no loss.
Fig. 3. Example of the experimental testbed, in which a round-robin scheduler
is used for splitting a UDP packet sequence over two heterogeneous paths to
a multilinked client.
This testbed was used in all the experiments described in the
following sections. The multilinked laptop is stationary and it
has access to two WLAN networks NW LAN1 and NW LAN2 ,
and an HSDPA network NHSDP A .
B. Packet Reorder Metrics
Various metrics have been defined for getting a notion of
the amount of reordering that occurs in a packet sequence.
In RFC 4737 [14], several metrics for packet reordering
are defined, each one focusing on certain application-specific
properties. For instance, packet reordering can be expressed as
the number of packets that arrived late (i.e., a packet is late,
if another packet with larger sequence number had previously
arrived). For expressing the percentage of packets that arrived
late, we will use the Type-P-Reordered-Ratio-Stream metric
defined in RFC 4737 and denote the metric as MT P RRS .
If reordering occurs to only a very small fraction of all
packets, MT P RRS is an intuitive metric for getting a sense of
Fig. 4. An example Reorder Density, obtained at a bitrate of 2.0 Mbit/s,
splitting packets in a round-robin fashion over an HSDPA and WLAN link.
For illustration purposes, the y-axis is in logarithmic scale.
Although MRD provides an exact summary of the reordering in a packet sequence, it is less intuitive than MT P RRS and
not suitable for comparing the disorder of two independently
measured sequences. A density function does not give a
clear indication of the system (e.g., scheduling solutions) that
performs the best. Thus, a singleton metric, called Reorder
Entropy, has been formulated in [12]. The metric MRE
expresses the total disorder of a packet sequence, including
both the fraction of displaced packets as well as the degree
to which the packets are displaced. The total Reorder Entropy
MRE is directly derived from MRD and defined as:
MRE = −
X
MRD (i) ∗ ln(MRD (i))
(1)
i∈D
The set D contains all existing displacements in MRD . It
will be shown how MRE grows with increasing bitrate.
BC = 0.32 Mbit/s. In a practical experiment, such as ours, BC
is not sharply defined because of delay variations. Therefore,
with an increasing bitrate, a gradual increase in reordering can
be observed. From the point where almost every packet on the
slow link is passed by a packet with higher sequence number
on the fast link, MT P RRS remains close to 50 %.
As opposed to the previously mentioned related work [10],
[11], this experiment confirms that in a multilink scenario,
the ratio of late packets goes far beyond the stated 0.01 % to
1.65 % and quickly approaches 50 % for bitrates greater than
the critical bitrate BC .
V. A NALYSIS OF PACKET L ATENESS
This section presents results obtained using the testbed
mentioned in Section IV-A and the packet reorder metrics
introduced in Section IV-B.
Figure 5 depicts the ratio of late packets according to the
metric MT P RRS in two sets of measurements. The two curves
were obtained by scheduling packets over NHSDP A and both
of the two available WLAN networks. In this experiment, the
average delay over NHSDP A was 80 ms, the average delay
over NW LAN1 was 42 ms, and for NW LAN2 , it was 44 ms.
Fig. 5. The metric Type-P-Reordered-Ratio-Stream expresses the percentage
of packets in a sequence that arrives late. For different WLAN networks, the
ratio of late packets is alike.
The vertical line located at roughly 0.32 Mbit/s indicates the
critical bitrate BC at which packets of the WLAN link start
to “overtake” packets sent over HSDPA, therefore causing the
latter to arrive late. Ideally, when each packet travels with a
constant delay, MT P RRS would jump from 0 % to 50 % when
the bitrate surpasses BC . If d1 and d2 denote the average
transmission delays over two different access networks and p
designates the packet size in bits, then the critical bitrate BC ,
at which late packets are expected, is defined as:
BC = p/(|d1 − d2 |)
(2)
In the experiment described above, the average difference
in transmission delays |d1 − d2 | is 37 ms, and the total packet
size used was 1480 bytes, which leads to a critical bitrate of
VI. M ULTILINK PACKET R EORDER R EDUCTION
There are several methods of reducing packet reordering at
the receiver. This section discusses two methods, a receiverand a sender-side approach, and presents experimental results
to show their gain.
A. Receiver-side Reorder Buffer
On the receiver side, packet reordering can be reduced
through the use of a buffer. Every packet that has a sequence
number greater than the currently expected sequence number
is put into a reorder buffer. If more out-of-order packets arrive
than fit into the buffer, the packet in the buffer with the
lowest sequence number (the most excessively delayed packet)
will be discarded. From an application’s point of view, such
discarded packets are lost (another method, not treated here,
would be to remove packets from the reorder buffer once a
certain application-specific delay has expired).
When a lot of reordering happens, the reorder buffer will
be heavily occupied, while in a scenario of permanent inorder delivery, the reorder buffer occupancy would amount
to zero. Figures 6 and 7 illustrate that for the combined use
of the two heterogeneous links NW LAN1 and NHSDP A , the
required buffer size increases with increasing bitrate and that,
if a certain number of discarded packets can be tolerated, the
buffer size can be reduced.
In order to capture the workload of such a reorder buffer,
RFC 5236 also defines the Reorder Buffer-Occupancy Density
metric, denoted here as MRBD . Figure 6 shows the results
of a reorder buffer implemented in our testbed as well as
the MRBD metric values obtained from it. For illustrative
purposes, a small buffer that can hold 4 packets was used.
When the bitrate was set to a low value, such as 0.1 Mbit/s, no
reordering occurred, and the reorder buffer was never utilized.
With a higher bitrate, such as 1.0 Mbit/s, packets experienced
reordering, and the buffer was occupied by 2 packets in 45 %
of the time and by an overall maximum of 4 packets in 5 % of
the time. If the buffer had been allocated to fit only 3 packets,
5 % of the received packets would have been discarded.
If the discard of packets is not allowed, the required size of
a reorder buffer quickly grows with an increasing bitrate (see
Fig. 7). In other words, the buffer size would be as large as the
most displaced packet in the corresponding MRD histogram.
On the other hand, applications that are resilient to a certain
degree of packet loss require a smaller buffer. In any case,
Fig. 6. The Reorder Buffer Occupancy Density metric gives a sense of
a receiver’s buffer requirements to reduce packet reordering. This example
shows the load of a buffer with a capacity of 4 packets at different bitrates. The
sum of each curve amounts to 100 %, indicating no packets were discarded.
packets with very high displacements might be delayed so
much that they are useless to the application.
Fig. 7. Receiver-side buffer requirements depend on the bitrate used and on
the number of packet discards that can be tolerated by the application.
Client-side buffers are not the only method to reduce the
occurrence of packet reordering. As shown in the following
section, the sender, or an intermediary proxy, is also able to
impact the total disorder in a packet sequence.
B. Sender-side Multilink Packet Scheduler
Packet reordering can also be mitigated by scheduling
packets in a more sophisticated way over the available network
interfaces than the simple round-robin assignment used in the
experiments discussed so far. The sender has two degrees of
freedom in deciding where to send packets. First, a link can
be favoured according to its bandwidth, so that more packets
travel over a fast link than over a slow one, and second,
packets can be sent at different points in time in the attempt
to compensate for heterogeneous delays.
To get a deeper insight into the heterogeneity of multiple
wireless links, we design a static scheduler. If the heterogeneity shows little variation, a static scheduler built on the average
delay and throughput should work close to perfect. An outline
of a static scheduler is given in Algorithm 1. In order to
guarantee a constant bitrate, packets are sent in equal time
intervals, and in each step the scheduler selects the network
interface to be used. The scheduler needs to have precognition
of the available interfaces and their estimated delay d1 and d2 .
The main idea of the scheduler is to delay sending packets
over the low-delay NW LAN1 link by putting them into a FIFO
queue. According to the delay difference |d1 − d2 | of both
interfaces and the bitrate used, it is calculated how large the
FIFO queue must be to delay the packets over NW LAN1 in a
way that they should experience a delay similar to the packets
sent over NHSDP A .
Algorithm 1 Static Scheduler using a FIFO Queue
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
Time interval [ms] TI = (bitrate∗106 )/(packetsize∗8) ∗ 1000;
Allocate queue Q of size q = b|d1 − d2 | /TI c;
for sequence number s = 1 to n do
loop
interface i = getInterfaceBySeqno(s);
if (i == NHSDP A ) then
exit loop;
else if (i == NW LAN1 ) then
Q.put(s);
if Q.full() then
Q.get(s);
exit loop;
else
s + +;
end if
end if
end loop
send sequence number s on interface i;
wait(TI );
end for
The function getInterfaceBySeqno() used in Algorithm 1
decides to which of the receiver’s network interfaces a packet
is sent. The same function is also used at the receiver,
which makes it possible to determine on which interface a
lost packet should have arrived. The round-robin version of
getInterfaceBySeqno() is displayed in Algorithm 2.
Algorithm 2 “Round-Robin” getInterfaceBySeqno(s)
1: if (s modulo 2 == 0) then
2:
return NHSDP A ;
3: else if (s modulo 2 == 1) then
4:
return NW LAN1 ;
5: end if
Figure 8 shows how the total packet disorder reduces when
the static scheduler is used. Compared to the purely roundrobin packet allocation, MRE decreases on average by 38 %
when using a queue that buffers packets before sending them
on the fast link.
dynamic wireless environment. Therefore, our future research
will include the design of packet scheduling mechanisms that
dynamically adapt to the current link characteristics based on
feedback from the receiver.
Furthermore, we will analyze the impact of multiple links
on packet reordering in TCP connections. We plan to implement a proxy solution without the requirement of server-side
modifications, enhancing the research done in [16]. Tunneling
mechanisms will be applied to transparently split traffic of the
same TCP stream over the proxy to all available client-side
interfaces.
VIII. ACKNOWLEDGEMENTS
Fig. 8. Comparison of the Reorder Entropy MRE for a purely round-robin
versus a static scheduler that is based on average link estimates (assuming
WLAN vs. HSDPA: delays = 40:80 ms, bitrate ratio = 3:1).
The authors would like to express their gratitude to Haakon
Riiser from Netview Technology AS (www.netview.no) for his
support and for providing us with a research license of the
Sigma Network Analyzer. We also thank Erlend Viddal from
Lividi, Eunah Kim from ETRI, Peter Kaspar from ETH Zurich,
and Matthias Wille from Avaloq for their valuable feedback.
R EFERENCES
The main problem of static schedulers is that the delay may
vary over time and that they are dependent on the bitrate
used. The closer the bitrate approaches link saturation, the
more will the experienced delays increase. Therefore, a static
scheduler based on fixed delay estimates cannot be expected
to perfectly solve packet reordering in an environment with
dynamic heterogeneity as has been demonstrated in wireless
networks. The design of a scheduler that adaptively adjusts to
the link delays is indispensable.
VII. C ONCLUSIONS AND F UTURE W ORK
Through the aggregation of multiple concurrent links, it is
possible to achieve increased throughput, but at the cost of
high packet reordering at the receiver. Recent studies on multihoming and multipath routing have focused either on highspeed, wide-area networks or ignored the severe heterogeneity
of wireless links.
In contrast, our work focuses on dividing packet sequences
to hosts over multiple last-hop wireless networks. For this
scenario, we have demonstrated with practical experiments
that the dynamics and heterogeneity of multiple wireless links
causes reordering of IP packets that clearly exceeds the one
observed for common Internet connections.
By using sufficiently large buffers, packet reordering can be
avoided. However, for devices with high resource constraints,
the workload of using large buffers is expensive. Instead
of discarding many packets that arrive out of order at a
receiving device, it is also possible to reduce packet reordering
by intelligently scheduling packets on the sender-side. Our
experimental results show that a static sender-side scheduler,
that has knowledge about the average delay and throughput of
the links, is able to reduce client-side buffering by 38 % on
average.
However, a static scheduler that is guided by average
delay and throughput estimates works inaccurately in a highly
[1] L. Golubchik, J. C. S. Lui, T. F. Tung, A. L. H. Chow, W.-J. Lee,
G. Franceschinis, and C. Anglano, “Multi-path continuous media streaming: what are the benefits?” Perform. Eval., 2002.
[2] C. Hopps, “Analysis of an equal-cost multi-path algorithm,” IETF RFC
2992, November 2000.
[3] D. S. Phatak, T. Goff, and J. Plusquellic, “IP-in-IP tunneling to enable
the simultaneous use of multiple IP interfaces for network level connection striping,” Computer Networks, 2003.
[4] J. R. Iyengar, P. D. Amer, and R. Stewart, “Concurrent multipath
transfer using SCTP multihoming over independent end-to-end paths,”
IEEE/ACM Trans. Netw., vol. 14, no. 5, 2006.
[5] B. Wang, W. Wei, Z. Guo, and D. Towsley, “Multipath live streaming
via TCP: scheme, performance and benefits,” in ACM CoNEXT. New
York, NY, USA: ACM, 2007, pp. 1–12.
[6] I. Gapanova, “Concurrent multipath transfer for heterogeneous networks,” Master’s thesis, TU Braunschweig, 2008.
[7] C.-M. Huang and C.-H. Tsai, “WiMP-SCTP: Multi-path transmission using stream control transmission protocol (SCTP) in wireless networks,”
in AINA Workshops (1), 2007, pp. 209–214.
[8] J. Liao, J. Wang, and X. Zhu, “cmpSCTP: An extension of SCTP to
support concurrent multi-path transfer,” ICC, 2008.
[9] N. M. Piratla and A. P. Jayasumana, “Reordering of packets due to
multipath forwarding – an analysis,” in International Conference on
Communications (ICC), 2006.
[10] S. Jaiswal, G. Iannaccone, C. Diot, J. Kurose, and D. Towsley, “Measurement and classification of out-of-sequence packets in a Tier-1 IP
backbone,” ACM SIGCOMM Internet Measurement Workshop, 2002.
[11] L. Gharai, C. Perkins, and T. Lehman, “Packet reordering, high speed
networks and transport protocol performance,” 13th International Conference on Computer Communications and Networks (ICCCN), pp. 73–
78, October 2004.
[12] B. Ye, A. P. Jayasumana, and N. M. Piratla, “On monitoring of end-toend packet reordering over the Internet,” in ICNS ’06: Proceedings of
the International conference on Networking and Services. Washington,
DC, USA: IEEE Computer Society, 2006, p. 3.
[13] N. M. Piratla and A. P. Jayasumana, “Metrics for packet reordering
- a comparative analysis,” in International Journal of Communication
Systems, 2007.
[14] A. Morton, L. Ciavattone, G. Ramachandran, S. Shalunov, and J. Perser,
“Packet reordering metrics,” IETF RFC 4737, November 2006.
[15] A. Jayasumana, N. Piratla, T. Banka, A. Bare, and R. Whitner, “Improved packet reordering metrics,” IETF RFC 5236, June 2008.
[16] K. Chebrolu, B. Raman, and R. R. Rao, “A network layer approach to
enable tcp over multiple interfaces,” Wireless Networks, vol. 11, no. 5,
pp. 637–650, 2005.